The repository contains the 4th place solution to Zindi's low-resource automatic speech recognition competition. The goal of the competition was to build a system that allows people to navigate public transport without being able to read or speak French - the language heavily used in the naming of local landmarks.
The solution consists of a finetuned wav2vec2-xlsr model, with predictions decoded using an n-gram language model and a nearest-neighbour search. Read the short description of the solution here
The script was tested on a high-RAM P100 Google Colab session. The full run should take about 6 hours.
- Place files in
/content/zindi-ai4d-wolof - Place competition data in
/content/zindi-ai4d-wolof/data - Run
run.sh
The output file is /content/zindi-ai4d-wolof/submission.csv